Scalable oversight is often discussed as the route to controlling superintelligent AI, but none of the researchers who build these methods asserts that it can, and the current evidence weighs against the claim without settling it.

Scalable oversight (debate, amplification, weak-to-strong training, peer prediction, chains of overseers) is commonly presented as the technical answer to supervising AI systems more capable than their supervisors, and discussion often runs on to superintelligent systems. The graph's assessment of the claim that such methods can reliably control superintelligent systems is doubtful, with a credence of about 0.10.

The record shows that the people with the strongest interest in the methods working do not make the claim. The paper that proposed debate called its practical success an empirical question. Google DeepMind's 2025 safety plan, which makes amplified oversight central, sets superintelligence aside as future work and says its approach does not need to be robust to arbitrarily capable systems. The debate safety case concerns an AI research agent, not a superintelligence. The assessment weighs this restriction of scope as an admission against interest about what is established, not as proof that the claim is false.

On the evidence, the assessment finds four things the claim would need together: an oversight signal that stays accurate at a very large capability gap, robustness of that signal to a system optimising against it, no build-up of error along a chain of overseers, and a route from an accurate signal to safe behaviour where the signal is not applied. It finds partial and conditional support for the first for particular methods, mixed evidence on the second with adaptive attacks succeeding where they have been tried, and no experiments on the third or fourth. The premises it rests on are assessed in their own right: that protocols remain effective against deliberately deceptive systems is doubtful, that the assumptions of the debate proofs hold for real tasks is doubtful, that errors do not compound across generations is doubtful, and that debate lets weaker judges supervise stronger debaters is inconclusive. The two best-supported premises cut in different directions: weak-to-strong outperformance is supported but concerns accuracy on tasks with known answers, and oversight success tending to fall as the capability gap grows is supported, though the measured pattern is uneven and the widely quoted low success figures for chained oversight are extrapolations from fitted lines.

The claim is doubtful and not contradicted because it concerns what a class of methods can do, some methods do scale favourably, and the decisive experiments have not been run. The assessment names what would move it: oversight that holds as the gap widens on tasks without checkable answers, protocols that withstand systems trained to defeat them, and evidence that the debate assumptions hold outside mathematics-like tasks. Two premises about whether an accurate oversight signal carries over to deployed behaviour are recorded on the claim and not yet assessed.